Performance analysis of remote photoplethysmography deep filtering using long short-term memory neural network.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Remote photoplethysmography (rPPG) estimates heart rate using standard video cameras.
• Conventional filters (bandpass, wavelet) are used to reduce noise but leave residual alterations.
• Experienced observers can identify remaining signal artifacts after conventional filtering.

Purpose of the Study:

• To investigate the efficacy of a long short-term memory (LSTM) network for filtering remote photoplethysmography (rPPG) signals.
• To identify and mitigate residual noise artifacts in rPPG signals using a deep learning approach.
• To compare the performance of LSTM-based filtering against conventional methods.

Main Methods:

• A many-to-one and many-to-many LSTM network architecture was employed for rPPG signal filtering.
• Experiments were conducted using three public rPPG databases in intra-dataset and cross-dataset settings.
• The LSTM network was trained on approximately 45 minutes of rPPG signal data.

Main Results:

• The LSTM-based filter demonstrated superior performance compared to conventional filters in intra-dataset scenarios.
• On the VIPL database, the LSTM achieved a Mean Absolute Error (MAE) of 3.9 bpm, outperforming conventional filters (improved from 10.3 bpm to 7.7 bpm).
• Cross-dataset performance showed a dependency on the signal-to-noise ratio, with closer values between training and testing sets yielding better results.

Conclusions:

• LSTM-based filtering offers a significant improvement over traditional methods for rPPG signal processing.
• A relatively small dataset (approx. 45 minutes) is sufficient to train an effective LSTM deep-filter.
• The LSTM network provides a robust and efficient solution for enhancing the accuracy of heart rate estimation from rPPG signals.